Laplace noise on each bin count. Because the bins are disjoint, one
record touches exactly one of them: this is PARALLEL composition, so
the same epsilon covers every bin at once. Dividing the budget across
bins is the common error and wastes it entirely.
Usage
morie_dp_histogram(
x,
bins = 10,
epsilon = 1,
range_ = NULL,
seed = NULL,
nonneg = TRUE
)
Arguments
- x
values to bin.
- bins
bin count or explicit edges.
- epsilon
privacy budget.
- range_
optional c(min, max) for the binning range.
- seed
optional integer seed.
- nonneg
clamp negative counts to zero (post-processing, free).
Value
list with release, raw, true_counts,
edges, noise_scale.
Details
The sensitivity used is 2, which covers the add/remove-one neighbouring
relation where a record can move between bins.
References
McSherry, F. (2009). Privacy integrated queries.
SIGMOD, 19-30. (parallel composition)
Examples
set.seed(1)
morie_dp_histogram(rnorm(500), bins = 5, epsilon = 1)$noise_scale
#> [1] 2
